Dexter

[Verax @veraxlab]

Over a forty year career, the 1 percent fee on your retirement fund takes more money out of it than you ever put in.

Not a typo. The US Department of Labor prints a smaller version of the same sum in its own pamphlet, and the single line of arithmetic under it explains why a fee behaves nothing like the small skim you assume it is.

Their example. Twenty five thousand dollars, thirty five years, seven percent a year. One fund takes 0.5 percent, the other takes 1.5. The cheap one finishes at 226,556 dollars. The expensive one finishes at 162,846.

The gap is 63,710 dollars. That is 28 percent of the account, and more than twice what was put in at the start.

Here is the line of arithmetic that does it. A fee is not subtracted once. It is charged on your balance every year, so this year it also takes a cut of the growth it prevented last year, and the year before that. You do not compound at 7 percent. You compound at 7 minus the fee, for as long as you hold it.

One percent looks like a rounding error because we compare it to the return. Compare it to the decades instead.

Now the lever, and it costs you one afternoon.

Start with 20,000 dollars, add 600 a month, run it 40 years at 7 percent. In a fund charging 1 percent you finish with 1,414,044. In an index fund charging 0.03 percent you finish with 1,884,043.

The difference is 469,999 dollars. Across those same 40 years you deposited 308,000.

Do it tonight. Open your retirement account, find what you are actually holding, and look for the line called expense ratio. It is one number and it is public.

Most people cannot simply switch, and that is worth saying plainly. A 401k is a menu your employer picked, and the cheapest thing on it may still be 0.4 percent. Two moves work anyway. Take the cheapest index option on the menu, whatever it is. And the old 401k from a job you left is on nobody's menu, it can be rolled over, and that is usually where the 1 percent funds are quietly sitting.

Now the part almost nobody knows. Until 1980 the SEC held that a fund must not use your money to sell itself to other people. The industry bled assets through the seventies, lobbied, and in 1980 the SEC reversed its own position and adopted Rule 12b-1.

It lets a fund charge the current shareholders for the advertising that brings in the next ones. Capped at 1 percent a year. In 1980 it collected a few million dollars. By 2009 it was collecting 9.5 billion.

Go read your expense ratio tonight. Part of that number is the cost of the ad that recruited the person after you.

[Dexter @dexterscare]

@vexelxbt this was way too close lol

[Vexel @vexelxbt]

96.6 percent accuracy. That is the track record an Army sergeant named John Finley reported in 1884, after America's first experiment in tornado forecasting.

One critic ran one check. Writing "no tornado" every single day, on the same 2,803 forecasts, would have scored 98.2 percent. The best record in the country lost to doing nothing at all.

Wall Street would have given Finley a fund. Meteorology gave him an audit - and then made the audit a daily habit. Forecasts are graded against the next morning, thousands of times a year. When statisticians checked decades later, forecasters who said 70 percent were right about 7 times in 10. Almost no other profession can prove that about itself.

This reel is that machine at work in 1925: the Weather Bureau lifting instruments on giant kites, Navy seaplanes, the airships Shenandoah and Los Angeles. Full film, from the National Archives, digitized last year. Almost nobody has watched it.

Your results have a number too - how many attempts before they prove anything at all. It is one line of arithmetic. I worked it out for poker players, fund managers and the market itself in the article below. Most people who run it do not like their answer.

[Verax @veraxlab]

Three couples kissed at an airport. Two turned their heads to the right.

That's 67 percent. It also means nothing, and every trader with a good month is doing the same arithmetic.

An MIT statistician uses this example in the first lecture of his statistics course, and it's free on OpenCourseWare. A study in Nature watched 124 couples kiss. 80 turned right. Is that a real bias, or noise? Pure chance would give you 62. To say anything at all you need 72. To be 99.9 percent sure, you need 80.

Nobody feels that threshold. Three wins in a row feel like proof. They aren't even a data point.

A man in Basel spent twenty years on exactly this question. Jacob Bernoulli died in 1705, and the answer came out in 1713, printed by his nephew from the manuscript he left behind. His own example: an urn with 3,000 white pebbles and 2,000 black. How many draws before you can trust the ratio you see? His number was 25,550.

Every fund that shows you a three year track record, every guru with a hot quarter, is selling you 67 percent from three couples at an airport.

The rule that tells you when results start to mean something has been free for 313 years. The only thing that costs anything is skipping it.

[Vexel @vexelxbt]

Everyone knows what caused 2008. Poor people were handed mortgages they could never repay, the loans went bad, and the system came down with them.

An economist went and pulled the actual credit files. Not surveys, not a sample of loans. A large administrative panel of individual credit records, tracking the same people from 1999 to 2013.

The story does not survive contact with them.

The borrowers who got blamed were not the ones driving the collapse. While the crisis was happening, their share of foreclosures fell by half. And the group that actually drove the defaults is one almost nobody was angry at, because they never looked like victims of predatory lending at all.

They were landlords.

In the data they are defined without emotion: a real estate investor is any borrower holding two or more first mortgages at the same time. Second homes. Rentals. Flips.

For borrowers in the middle of the credit score distribution, the share of mortgage balances held by these investors climbed from about twenty percent to thirty-five between 2004 and 2007. When it broke, their share of foreclosures went from roughly twenty percent to about sixty.

Over the same collapse, the share of foreclosures coming from the worst quarter of credit scores fell from around seventy percent to thirty-five.

Read that again. In the crisis named after subprime borrowers, the subprime share of foreclosures halved.

The paper's own summary of the boom years is blunt. Credit growth between 2001 and 2007 was concentrated in the prime segment, and debt to high risk borrowers was virtually constant.

So how did everyone get it backwards for a decade?

Her answer is almost boring, which is what makes it useful. Earlier work looked at loans instead of following people. Borrowers who were young when the boom started took on more debt as they aged, the way people always do. Look at that in a snapshot and it reads as credit being pushed onto bad risks. Follow the same individuals through time and it reads as a life cycle.

One methodological choice, and an entire group carries the blame for a generation.

That is the part worth keeping, and it has nothing to do with mortgages. When you are shown who is responsible for something, the question is not whether the number is right. It is who got counted, over what stretch of time, and whether the same people are inside the number at the start and at the end.

The lecture runs forty-three minutes. In seven years, about nineteen hundred people have watched it.

[Dexter @dexterscare]

@vexelxbt amazing article broski

[Vexel @vexelxbt]

https://t.co/62wJFf5wYq

[Dexter @dexterscare]

@vexelxbt really interesting experiment

[Dexter @dexterscare]

@0xkemu The model didn’t improve. The measurement environment did. That distinction matters more than the headline.

[kemu @0xkemu]

THE MOST COMMON FAILURE IS NOT A MADE-UP ANSWER. IT IS A CONFIDENT NOTHING FOUND.

Two of the ten repos on that list say it, in their own files, in different words.

One calls it the single most common observed failure. More common than fabrication.

The other found that every failure in its retrieval layer looks identical from the outside. No results.

Every demo shows you the answer. None of them show you the question that quietly came back empty.

A made-up fact you can check. An empty answer you cannot - the thing that would prove it wrong is the thing that was never found. It just becomes something you believe your notes do not contain.

The fix is one sentence added to how you ask:

Before you tell me something is not there, show me how you looked. Paths listed, terms grepped, files opened. If you cannot show the search, say you did not search.

That works on any folder, with any agent, with none of these repos installed.

Now the numbers that should annoy people.

Same vault, same question, same build, run twice. With the search index working: 23 turns, 72.8 seconds. With the index deleted - all 1,267 notes, nothing to fall back on: 26 turns, 76.2 seconds.

Three more turns. 3.4 more seconds. The index makes it cheaper. It does not make it work.

Go ask yours something you know is in there, worded slightly wrong on purpose. If it comes back empty and names no path it opened, you cannot tell whether it looked. That is the whole problem, and it is one sentence away from fixed.

[Dexter @dexterscare]

@0xkemu Useful, bookmarked

[Vexel @vexelxbt]

The best investment fund in recorded history threw out all of its clients.

It returned 66 percent a year for thirty years. Then its founder decided outside money was a problem, closed every outside account, and never let anyone back in.

$100 in the US stock market in 1988 became $1,910 by 2018. The same $100 in Medallion became $398 million.

The man in this video built it. A mathematician and code breaker who never worked a day on Wall Street. His story starts with a small business in Colombia - when it paid off, he took the money and went looking for patterns.

The fund charges the worst fees in the business. 5 percent flat plus 44 percent of profits. His own employees pay them gladly, because 66 gross still comes out to 39 net, and 39 beats everything ever measured.

So why throw out the clients?

Medallion finds prices that are slightly wrong and corrects them before anyone notices. An edge like that has a fixed size. Every outside dollar squeezes into the same gap and makes the returns smaller for the dollars already in.

The fund did not close because it failed. It closed because it worked.

Anything that truly earns 66 percent a year has no reason to be offered to you.

So when an edge shows up in your feed with a price tag on it, ask the only question that matters.

If this works, why is there room for me?

[kemu @0xkemu]

TEN REPOS WANT WRITE ACCESS TO YOUR NOTES. ONE OF THEM PROTECTS A SENTENCE YOU TYPED YOURSELF.

Not the biggest one. The 44.3k repo at the top of that list gives the agent a create command with an overwrite flag and a JavaScript eval against the vault API, and never names one thing that is off limits. It is a CLI reference, not a safety layer. It is on the list anyway.

Every setup guide for these tools shows you the ingest. Nobody shows you the diff.

Clone all ten and read the command surface instead of the README. Three name nothing they will not touch. Four hold the agent back some other way - never delete, ask first, prompt on every action. Two fence off a folder. One goes inside the file:

<!-- human:start -->
your words
<!-- human:end -->

Two lines. Wrap what you wrote by hand, then tell the agent never to rewrite anything between the human markers. Nothing to install, no repo required, works in any vault. Source: references/write-rules.md in eugeniughelbur/obsidian-second-brain.

For a pass that cannot write at all, skip plan mode and take the tools away instead: --tools Read,Grep,Glob. Ask the model which tools it has under that flag and it will list Write. Ask it to write a file and it cannot. It reports its prior, not its runtime. Source: ARCHITECTURE.md in breferrari/obsidian-mind.

Capability is checked by trying, never by asking.

[Verax @veraxlab]

A Yale professor offered his class five dollars to pick the right number. The correct answer turned out to be the most popular one in the room. Every person who gave it lost, and that is the whole difference between being right about a price and getting paid for it.

Ben Polak, ECON 159, Yale. The entire course is online for nothing.

The rules were one line. Write down a whole number between 1 and 100. Whoever lands closest to two-thirds of the class average takes the money.

The correct answer is 1. Two-thirds of 100 is 67, so nothing above 67 can win. If nobody picks above 67, then nothing above 44 can win either. Run it far enough and every number burns off except 1. In that room, 1 was the most common answer written down.

The class average came in at about 13 and a third. Two-thirds of that is 9. The people who wrote 9 took the five dollars.

That 9 is not luck. It is a measurement. Start at 50, the number you write if you assume nothing about anyone. One step of thinking is 33. Two steps, 22. Three, 15. Four, 10. Yale stopped at four.

Richard Thaler ran the same contest on Financial Times readers in 1997. 1,382 entries, average 18.9, winner 13. He ran it again eighteen years later. 583 entries, average 17.3, winner 12.

Two decades apart, different people, same depth. Three steps, four at the outside. That is not a room being stupid. That is where human beings stop, and it has held every time anyone has bothered to measure it.

You have played this before. Every time you sold anything, a car, a flat, an old phone, you did not set the price at what it was worth to you. You set it at what you guessed a stranger would think. Nobody taught you that. You worked it out because it is the only version of correct that pays.

The other version is the expensive one. Being right early feels exactly like being wrong. Same waiting, same silence, same money sitting there doing nothing. The only thing that tells them apart afterwards is whether
the room showed up, and the room is three steps deep.

Which is the part left out of every service that sells you a mispriced asset. They are selling you the number 1. Provable, elegant, and worth nothing until enough strangers get there too. They might take a year. They might take nine. They might never come.

Yale gave the whole course away. The five dollars was the only scarce thing in that room.

[Dexter @dexterscare]

@veraxlab bookmarked

[Razex @0xRazex]

this is insane. patrick winston spent decades at mit teaching some of the smartest people in the world, and one of the most useful things he ever taught had almost nothing to do with artificial intelligence.

every january, he gave the same lecture: how to speak.

his point was simple. you can have the best idea in the room and still lose if nobody understands it.

people spend thousands on communication coaches learning versions of what winston taught in one hour for free. don't read from slides. keep visuals simple. remove clutter. connect with the audience early. finish with something they can actually remember.

he believed speaking wasn't a soft skill. it was a multiplier on everything else you knew.

that's why the smartest engineer doesn't always get the promotion. the best founder doesn't always raise the money. the strongest idea doesn't always win.

we like to believe good work speaks for itself. most of the time, it doesn't. someone has to make people understand why it matters.

winston recorded the lecture at mit in 2018. he died the following year.

millions of people have watched it since.

the information isn't the advantage anymore. everyone has access to it.

the advantage is being one of the few people who actually uses it the next time they enter a room.

[Dexter @dexterscare]

@veraxlab Bookmarked

[Verax @veraxlab]

Everyone starts with $100. The bet has positive expected value. After a hundred rounds the median player has 52 cents.

The article below runs a simulation and shows you the winners. Run the cleanest version of the same rule and look at everyone else.

One coin flip per round. Heads, your money grows 50 percent. Tails, it shrinks 40 percent. Fifty-fifty, forever. Work out the expected value and it is plus 5 percent per round. Every textbook tells you to take that bet.

Now actually run it.

- Expected pot after 100 rounds: $13,150
- Median player: $0.52
- 86 out of 100 people end with less than they started

Nothing was rigged. No fees, no crash, no villain, no head start. The average rises because a few paths go vertical and drag it up. The person in the middle is quietly wiped out. Both sentences are true, out of the same twelve lines.

The reason is that an average is taken across people, and your life is taken across time. For anything that grows by a
percentage, those two numbers are not the same number. Ole Peters gave a public lecture on exactly this at Gresham College in London. Gresham has been giving free lectures since 1597. It is on video, whole, for nothing.

So the tail is real. So is this: the expected return in the brochure describes a person who does not exist.

Bookmark it. The next time someone shows you an average return, ask what happened to the median.

[Vexel @vexelxbt]

90 percent confident. Right about 10 percent of the time.

That is what happened when a Yale class was asked to guess a range for the Earth's mass, wide enough that they were 90 percent sure the real number sat inside it. Almost all of them missed.

Same students, same method, asked to guess the world's population instead. This time they got it right about 80 percent of the time. One question they could judge. Two they could not, and nothing told them which was which.

That gap between how sure you feel and how right you are is not free. It has a price tag. Funeral insurance. Flight insurance. Insuring a diamond ring against loss. Every one of those products is priced for a brain that overweighs a rare, vivid risk and ignores the boring one that actually drains a bank account over time.

Yale filmed the whole lecture and put it online for free. Nobody profits from you watching it. They profit from you never connecting it to your own decisions.

[Verax @veraxlab]

A single study convinced an entire field that poverty steals ten points of IQ from anyone experiencing it. When other scientists went looking for that exact effect, almost none of them could find it.

MIT still teaches the original finding as a landmark result. The same professor also teaches the part almost nobody repeats out loud: it doesn't replicate.

The original study put shoppers in a mall in Trenton, New Jersey, and asked them to think about their finances before an IQ-style test. Low-income shoppers lost about 10 points. Wealthier shoppers lost nothing.

A second version tested real farmers in Tamil Nadu, India, before and after harvest. Same people, same brains.

Broke before harvest, they scored worse. Flush with cash after, they scored better.

The finding anchored an entire influential book. Poverty does not just take your money, it takes a slice of your mind. It got cited everywhere.

Then other researchers went looking for it again. No difference around paydays among US households. No difference before and after harvest among farmers in Zambia.

The most viral number in behavioral economics barely survived contact with anyone else's data.

Frank Schilbach teaches this exact history to MIT graduate students. The finding and the failed replications, both, in a lecture that has been free online since 2021.

Almost none of the people who repeat that ten-point number online have ever heard him say the second half of the sentence.

Poverty is still bad for the mind. Cash transfers reliably make people happier. Losing a job reliably makes people worse off. Those parts hold up.

It's the one famous number everyone quotes that might not.